For the past few years, much of the conversation about generative AI in higher education has focused on one problem:
Students can use AI to do the work.
And that matters.
If an assessment can be completed convincingly by AI without the student understanding the content, then the assessment may need to change. More oral questioning, demonstration, application, viva-style defence, practical work, or follow-up questioning may make the learner’s own understanding more visible.
But assessment is only one part of education.
I increasingly think we need a second conversation:
When should we actually encourage students to use AI?
Because protecting assessment and supporting learning are not the same problem.
Research on generative AI in education increasingly reflects this tension. Large language models can provide explanations, feedback and personalised support, but their value depends heavily on how they are used and how the learning activity is designed (Kasneci et al., 2023).
The question is not simply: “Did you use AI?”
I think a more useful question is:
What did you use AI for?
There is a major difference between asking:
“Write this assignment for me.”
and asking:
“I don’t understand preload. Explain it at first-year physiotherapy level, then ask me three questions to check whether I understand it.”
Both involve AI.
But educationally, they are doing very different things.
The first may remove the thinking.
The second may create more opportunities for it.
That distinction should probably sit at the centre of how we teach students to use AI.
A simple principle: keep the important thinking with the learner
The principle I keep coming back to is:
AI should make learning easier without making thinking optional.
That means reducing unnecessary barriers, providing timely explanations and giving feedback — while preserving the mental work we actually want students to practise.
There is evidence that this distinction matters.
In a large field experiment with nearly 1,000 secondary-school mathematics students, Bastani and colleagues compared a relatively unrestricted GPT interface with a more carefully designed GPT tutor. Both improved performance while students had access to AI, but students using the unrestricted version later performed worse than controls when the AI was removed. The more structured tutor largely avoided that negative effect (Bastani et al., 2025).
That is an important distinction for us as educators:
Performance with AI is not automatically evidence of learning without AI.
So how should we use it?
I would use a fairly simple sequence.
- Student first. Ask the learner to attempt the problem, retrieve what they know, make a decision or explain their current understanding before consulting AI.
- AI second. Use AI for explanation, hints, feedback, alternative examples or questioning rather than simply providing the completed answer.
- Student explains. Ask the learner to explain the idea back, justify the answer, or describe why they changed their reasoning.
- Verify. Require the learner to check AI output against lecture material, textbooks, clinical guidance or other appropriate sources.
- Remove the support. End with something the learner has to do independently: a new case, a short viva, a practical demonstration, or an explanation without AI.
For me, that sequence captures the difference between AI as a crutch and AI as a scaffold.
Step 1: let the student struggle a little
I do not mean that students should be left confused indefinitely.
But I also do not think AI should always arrive before the learner has even attempted the problem.
There is educational value in trying.
Retrieval practice is well established as a way of improving later retention, rather than merely measuring memory (Roediger & Karpicke, 2006).
There is also a broader idea behind approaches such as productive failure: attempting a difficult problem before receiving full instruction can sometimes prepare learners to understand the later explanation more deeply (Kapur & Bielaczyc, 2012).
So instead of:
“Ask ChatGPT how cardiac output changes during exercise.”
I might ask:
“Before using AI, write down what you think happens to heart rate, stroke volume and cardiac output during exercise, and explain why.”
Then AI enters the conversation.
Now the learner has something to compare, defend or revise.
Step 2: tell AI not to give the answer
This may be one of the most practical things we can teach students.
Instead of asking:
“What is the answer?”
try:
“Do not tell me the answer. Ask me one question at a time until I work it out.”
Or:
“Here is my reasoning. Challenge it, but do not rewrite it for me.”
Or:
“Give me one hint only.”
Or:
“Explain where my reasoning goes wrong, then give me a similar problem to solve independently.”
The interesting thing is that the same technology that can remove thinking can also be instructed to demand thinking.
That difference is not trivial.
A randomised controlled study in an undergraduate physics course found that a carefully designed AI tutor, built around pedagogical principles and structured prompts, produced greater learning gains than the comparison active-learning classroom condition in that particular setting (Kestin et al., 2025).
The authors emphasised that the tutor was deliberately designed around active engagement, scaffolding, targeted feedback and self-pacing; it was not simply an unrestricted chatbot.
That distinction is important.
It may not be AI versus no AI.
It may be:
poorly designed AI use versus pedagogically designed AI use.
Step 3: make the learner explain
This is where I think we can make AI much more educationally useful.
Suppose a student asks AI about elbow goniometry.
Instead of allowing:
“Where should the axis go?”
we could require:
“I think the axis should go here because… Tell me whether my reasoning is sound.”
Or after receiving help:
“Now explain back to me why the lateral epicondyle is used.”
The self-explanation literature gives us good reason to value this kind of activity. Chi and colleagues found that prompting learners to explain material to themselves supported deeper understanding compared with simply rereading it (Chi et al., 1994).
So perhaps one of the simplest AI-design rules is:
Never let the conversation end with the AI’s explanation.
Make the student respond.
Step 4: use AI where individualisation actually helps
This is where I think AI has a genuine advantage.
In a class of many students, I cannot always provide a different explanation to every learner at exactly the moment they need it.
One student may need simpler terminology.
Another may need a clinical example.
Another may need a diagram described in words.
Another may need to be challenged with a harder case.
Another may simply need to hear the explanation again without feeling embarrassed about asking for the third time.
AI can potentially provide that kind of on-demand support.
Kasneci and colleagues argued that large language models have potential for personalised learning and support, while also stressing important limitations including inaccuracies, bias and over-reliance (Kasneci et al., 2023).
The opportunity, then, is not simply that AI gives answers quickly.
It is that it can adapt the route to understanding.
Step 5: always end with independence
This is probably the most important safeguard.
If students learn with AI, at some point I still want to know:
Can you do this without it?
That might mean:
- answering a new question;
- explaining the concept aloud;
- completing a practical task;
- analysing a different case;
- defending a decision;
- completing a short closed-AI quiz.
This is where assessment and learning reconnect.
AI may help the student get there.
But ultimately, the learner still has to own the knowledge.
What this might look like in physiotherapy
Imagine a shoulder-pain clinical reasoning tutorial.
Instead of saying:
“Use ChatGPT to solve this case.”
I might structure the activity like this:
First: identify three initial hypotheses yourself.
Then: ask AI to challenge your hypotheses.
Then: explain which findings support or weaken each one.
Then: ask AI for one alternative hypothesis you may have missed.
Finally: close the AI and make your own final clinical decision.
Now AI is not doing the clinical reasoning.
It is creating another layer of questioning around the student’s reasoning.
The same idea could work in practical skills.
The learner places a virtual goniometer.
The simulation determines that the axis is too distal.
Instead of revealing the answer, the AI tutor asks:
“Which anatomical landmark should act as the fulcrum for this measurement?”
If the learner struggles again, the support becomes more explicit.
The computer detects the error.
The AI scaffolds the learning.
The student still has to make the correction.
A useful distinction: friction versus unnecessary friction
I think this may be one of the most useful ways to think about AI.
Some difficulty is educationally valuable.
Remembering.
Reasoning.
Explaining.
Making a decision.
Recognising that you were wrong.
Those are forms of productive friction.
Other difficulty may add little:
spending twenty minutes trying to understand one badly written sentence;
being too embarrassed to ask for another explanation;
waiting a week for feedback on a simple misconception;
not knowing where to begin because the task feels overwhelming.
AI may be very good at reducing that unnecessary friction.
The challenge is to remove the second without accidentally removing the first.
Perhaps we need two AI policies, not one
I increasingly think our approach should distinguish between two contexts.
When we are assessing
The question is:
What can this student do independently?
If unrestricted AI undermines that judgement, then restrict it or redesign the assessment.
Ask students to explain.
Demonstrate.
Apply.
Defend.
Respond.
When students are learning
The question changes:
What support will help this student eventually perform independently?
Here, AI may be useful.
It can explain.
Ask questions.
Generate practice.
Provide hints.
Challenge reasoning.
Give feedback.
Offer another example.
But it should be designed so that the learner does not disappear from the learning process.
The guide I would give students
If I had to reduce all of this to one rule for students, it would be:
Try first. Use AI to help you think. Explain what you learned. Check it. Then try again without AI.
That feels much more useful than simply saying:
“Do not use AI.”
And it is also more useful than:
“AI is allowed.”
Both of those statements are too simple.
What matters is what the tool is doing to the learner’s thinking.
Maybe the goal is not less AI
I am becoming less interested in asking whether students used AI.
I am more interested in asking:
Where did the thinking happen?
If AI generated the answer and the student submitted it, then the important thinking may have been outsourced.
But if the learner attempted the problem, explained their reasoning, used AI to receive a hint, revised their answer, checked the information and then solved a new problem independently, AI may actually have increased the amount of meaningful learning activity.
So perhaps the goal is not:
Use less AI.
Perhaps the better principle is:
Use AI in ways that leave the important thinking with the learner.
That seems, to me, a better place for the next phase of the conversation.
References
Bastani, H., Bastani, O., Sungu, A., Ge, H., Kabakcı, Ö., & Mariman, R. (2025). Generative AI without guardrails can harm learning: Evidence from high school mathematics. Proceedings of the National Academy of Sciences, 122(26), e2422633122. https://doi.org/10.1073/pnas.2422633122
Chi, M. T. H., De Leeuw, N., Chiu, M.-H., & Lavancher, C. (1994). Eliciting self-explanations improves understanding. Cognitive Science, 18(3), 439–477. https://doi.org/10.1207/s15516709cog1803_3
Kapur, M., & Bielaczyc, K. (2012). Designing for productive failure. Journal of the Learning Sciences, 21(1), 45–83. https://doi.org/10.1080/10508406.2011.591717
Kasneci, E., Seßler, K., Küchemann, S., et al. (2023). ChatGPT for good? On opportunities and challenges of large language models for education. Learning and Individual Differences, 103, 102274. https://doi.org/10.1016/j.lindif.2023.102274
Kestin, G., Miller, K., Klales, A., Milbourne, T., & Ponti, G. (2025). AI tutoring outperforms in-class active learning: An RCT introducing a novel research-based design in an authentic educational setting. Scientific Reports, 15, 17458. https://doi.org/10.1038/s41598-025-97652-6
Roediger, H. L., III, & Karpicke, J. D. (2006). Test-enhanced learning: Taking memory tests improves long-term retention. Psychological Science, 17(3), 249–255. https://doi.org/10.1111/j.1467-9280.2006.01693.x